Foundations 02 5 min Updated Sep 20, 2026
Jev, chat models, and code — who owns what
Code owns control flow and side effects; Jev owns narrow semantic judgments; LLMs write prose when you need text. Prefer workflows + gates over open agent loops.
A durable split from TypeSafe’s how-to-build guidance:
| Owner | Responsibility |
|---|---|
| Code | Control flow, side effects, retries, permissions, composing answers |
| Jev (System One) | Narrow, typed semantic judgments over state |
| LLM | Prose, long-form reasoning, open-ended generation when the product needs text |
Anti-pattern
Do not default to an agent while loop that re-prompts an LLM for every gate when a workflow + Jev confidence gates would suffice. Open loops burn tokens and blur who decided what.
Practical rule
If the output must be a label, score, or yes/no probability that code will enforce, start with Jev. If the user needs a paragraph, draft with an LLM after (or beside) the decision—not instead of it.